Learning essential features and forming simple representations of electroencephalography (EEG) signals are dificult problems. Variational autoencoders (VAEs) can be used with EEG signals to learn the salient features of EEG data. But explainability should disclose knowledge of how the model makes its decision. The key contribution of this research is the combining of known components in a pipeline that allows us to give meaningful visualisations that help us understand which component of latent space is responsible for capturing which region of brain activation in EEG topographic maps. The results reveal that each component in the latent space contributes to capturing at least two generating factors in topographic maps. This pipeline can be used to produce EEG topographic maps of any scale. Furthermore, assist us in understanding each component of latent space responsible for activating a portion of the brain.
Electroencephalography (EEG) is a method of recording brain activity (electrical potentials)
using electrodes placed on the scalp [
Traditional Autoencoders (AE) aim to learn prominent latent representations from unlabeled
input while ignoring irrelevant features. Variational Autoencoders (VAEs) was recently proposed
as an efective extension of AEs, for modeling a data’s probability distribution and learning a
latent space, usually of a lower dimension. It is ideal for unsupervised learning to understand the
impact and importance of each latent component for capturing the number of true generative
factors. VAE-based latent space analysis and decoding of EEG signals are important since they
can precisely define and determine the latent relevant features [
In this study, if CNN-VAE is trained with spatially preserved EEG topographic maps and its interpretation of the learned latent space representation provides the knowledge of how well each component of the latent representation contributes to capturing the number of true generative factors in spatially preserving EEG topographic maps via visual plausibility. The detailed design of this research is illustrated in figure 1.
The DEAP dataset was chosen because it contains multi-channel EEG recordings with 32 participants who watched 40 one-minute music video clips and tasks [18]. These EEG signals were transformed into 40 × 40 interpolated topographic maps that preserve spatial information about brain activation [19], a similar experiment was carried out in another investigation [20], as illustrated in (figure 1, B). Following the creation of the topographic maps, a Convolutional Variational Autoencoder (CNN-VAE) is built. The encoder network of CNN-VAE takes a 40 × 40 tensor (as seen in figure 1, C) and defines the approximate posterior distribution ( | ). The CNN-VAE decoder is a generative network that takes a latent space as input and returns the reconstructed EEG topographic maps. The architecture (figure 1, C) is made up of four 2D convolutional layers, each followed by a max pooling layer to minimize the dimension of the feature maps. In each convolutional layer, ReLU is employed as the activation function. To avoid overfitting, an early stopping strategy with a patience value of ten epochs is used, which indicates that training is stopped if the validation loss does not improve for ten consecutive epochs. To examine the number of generative factors captured from each active latent component, the reconstructed EEG topographic map from the decoder of CNN-VAE with latent space representation of only one active component is passed as an input to the K-means algorithm. In terms of how well samples are clustered with other samples that are similar to each other, the silhouette score is used to evaluate the quality of clusters generated using clustering methods such as K-Means. The reconstruction capacity of this model is evaluated by Structural Similarity Index (SSIM), Mean Squared Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE) derived for the reconstructed topographic maps.
The reconstruction capacity of the CNN-VAE model and its explainability is described with two diferent scenarios because it contains two decoder networks. One trained with all latent components, whereas the other trained with only one active latent at a time. In the second scenario, interpreting the disentangled representation of CNN-VAE, where its decoder network is trained only with one latent component alternatively and the remaining 24 components are set to zeros to examine the impact of each component for generating the patterns in EEG topo maps. The results show that each component contributes diferently to capturing the generating aspects in topo maps. The empirical experiment was carried out using test data, with 10 samples chosen at random to assess the impact of each latent component in capturing the number of fundamental generative elements in spatially preserving EEG topographic maps. In order to analyse the results Using visual plausibility, ten images of test data and reconstructed images with active latent space components 0 are plotted, with findings clearly suggesting that each component is learning two to three patterns from those EEG topographic maps. These explanations can be used to gain the trust of stakeholders by demonstrating the visual plausibility of each latent component in capturing the generative components in EEG topographic maps.
Research on the interpretation of disentangled representations of VAE trained with spatially preserving EEG topographic maps is currently limited. A CNN-VAE decoder network is trained with alternatively one active latent component, and the remaining component is set to zero because the mean value is close to zero in the distribution learned from each latent component. The results with visual plausibility show that each component contributes diferently to capturing and generating aspects in topo maps. Hence, this pipeline helps us understand each component of latent space responsible for activating a part of the brain region. Future studies will include understanding the decision of CNN-VAE through the interpretation of its latent space via clustering and visual plausibility, taking into account the signal-to-noise ratio and correlation values across the input and output of the architecture. intelligence, volume 33, 2019, pp. 4610–4617. [17] V. Prasad, D. Das, B. Bhowmick, Variational clustering: Leveraging variational autoencoders for image clustering, in: 2020 international joint conference on neural networks (IJCNN), IEEE, 2020, pp. 1–10. [18] S. Koelstra, C. Muhl, M. Soleymani, J.-S. Lee, A. Yazdani, T. Ebrahimi, T. Pun, A. Nijholt, I. Patras, Deap: A database for emotion analysis; using physiological signals, IEEE transactions on afective computing 3 (2011) 18–31. [19] T. Ahmed, L. Longo, Examining the size of the latent space of convolutional variational autoencoders trained with spectral topographic maps of eeg frequency bands, IEEE Access 10 (2022) 107575–107586. doi:10.1109/ACCESS.2022.3212777. [20] A. V. Chikkankod, L. Longo, On the dimensionality and utility of convolutional autoencoder’s latent space trained with topology-preserving spectral eeg head-maps, Machine Learning and Knowledge Extraction 4 (2022) 1042–1064.